Random Topics in Graph Theory
Random Topics in Graph Theory
I hereby declare that this report has been composed by me. I also
declare that all sources of information have been specifically ac-
knowledged and all quotations distinguished by quotation marks.
(signed) ...............................
Summary
The route taken for this project was initially aimed at learn-
ing the basics of graph theory and reading the proof of Kempe-
Heawood’s Five Colour Theorem, however with time the project
evolved to included more interesting aspects of graph theory such as
random graphs and percolations discussed in the two last chapters.
0
CHAPTER 1
Basic set-up
In order to study graph theory we first need to formally define graphs together with
their parameters and structures appearing within. In the opening chapter we intro-
duce the notion of a graph and look at some well-known classes of graphs which will
be used later. Next, we will define what is meant by a planar graph and offer a proof
of the useful Euler’s Formula.
X
deg(v) = 2q (1.1)
v∈V
X XX
i(v, e) = i(v, e) = 2q
v∈V, e∈E e∈E v∈V
The name of Lemma (1.1) comes from the fact that a handshake by default is being
exchanged between two people. Similarily, each edge of G is shared by 2 vertices.
The technique employed here is an example of double-counting, that is counting the
size of a set in 2 different ways and showing that the obtained expressions are equal.
Some important examples of graphs which we will use later on include:
2. Connectivity
Trees are a special class of graphs which due to their simple structure obey the
following principle
Theorem 1.1. A graph G on p vertices and q edges is a tree if and only if it is acyclic
and
p=q+1 (1.2)
Proof. The shortest odd length circuit has length 3, since there is no circuit of
length 1. But then it is an odd cycle of length 3. Assume that all odd circuits of size
at most r contain an odd length cycle, for some odd integer r ≥ 3. We must show
that an odd circuit of lenght r + 2 contains an odd cycle. Consider a circuit of the
form v0 , . . . , vr+2 , v0 of odd lenght r + 2. If it is a cycle then we are done. Assume it
2this can be shown by considering u − v, the longest path on non-trivial tree graph T . Since it
is the longest path, neither u nor v can be adjecent to some other vertex which is not on the path.
Also, u is adjecent to a vertex proceeding it on the path, same for v. Finally, neither u nor v are
adjecent to another vertex on the path u − v because a tree is acyclic.
Chapter 1 Page 5
is not a cycle and there exist positive integers m, k such that m < k and vm = vk . We
can consider two separate circuits v0 , . . . , vm , vk+1 , . . . , vr+2 , v0 and vm+1 , . . . , vk . The
sum of their lengths is odd, hence one of the circuits must be odd and of length less
than r. By the hypothesis this circuit contains an odd cycle, hence v0 , . . . , vr+2 , v0
contains an odd cycle.
Coming back to the connected graph G we can define an eqivalence relation on the
vertex set V by the rule: v0 ≡ vn ⇔ there is a walk from v0 to vn . Each subgraph
induced by the vertices in an equivalence class is called a component of G. Note that
the relation satisfies reflexivity since there exists a trivial vx − vx walk, symmetry
since the existance of vx − vy walk implies existance of vy − vx walk and transitivity
since the existance of walks vx − vz and vz − vy implies that there is vx − vy walk.
Armed with the means of this section we can now give an important characterization
of previously introduced bipartite graphs.
accomplished dividing the vertex set of H into two partitions of vertices according to
the parity of the path joining va with vertices of H, hence arriving at a bipartition of
H.
Now that we defined what is meant by connected graph G, one can wonder what
is the measure of how well connected G is. For instance we would think that Cn
belongs to weakly connected graphs, since removing any two edges would disconnect
it, while Kn would pass as a strongly connected graph with a minimum of n − 1
edges resulting in disconnection. We can introduce edge-connectivity κ(G) to be the
minimal number of edges which need to be removed from G in order to disconnect
it. Although knowing which edges to remove does not always prove to be easy, one
notices that in the worst case we can always achieve disconnection by isolating the
vertex of the smallest degree.
3. Planar graphs
Graphs come in a wide range of shapes and sizes and there is a special class of the so
called planar graphs. A planar graph is a graph that can be embedded in the plane,
that is it can be drawn on the plane in such a way that no edges cross each other.
A short survey of the graphs we met so far reveals that all the cycle graphs Cn are
trivially planar, while complete graphs Kn are not for n ≥ 5. As for the bipartite
graphs the story is more interesting, while some bipartite gaphs like K2,3 are planar,
anybody who has ever tried to solve the 3 Utilities Puzzle knows that K3,3 is not
planar4.
Since every planar graph divides the plane into a set of disjoint regions, we can
distinguish a new feature emerging within planar graphs. Any cycle which surrounds
a region forms a face fi . The portion of the graph lying outside of the graph G is
called the infinte face f0 . If {f0 , f1 , . . . , fn } is the set of all faces of a planar graph
G then (f0 ∪ . . . ∪ fn ) is the entire plane. We say that two faces are adjacent if they
share an edge e with the shortest cycle which bounds a face having length 3. If we
denote the number of faces by r and the number of edges by q, double counting yields
3r ≤ 2q (1.4)
4A classic result due to Kuratowski states a graph G is planar if and only if it contains neither
K5 nor K3,3 as a topological minor
Chapter 1 Page 7
Figure 4. Both K4 and K2,3 are planar while neither K3,3 nor K5 is.
which becomes strict if at least one of the regions is not triangular. Our next result
proves to be even more convinient because it elegantly relates the number of faces,
edges and vertices of a planar graph G.
Theorem 1.3. (Euler’s Formula) Let G be a connected planar graph, and let p, q and
r denote, respectively, the numbers of vertices, edges, and faces in a plane drawing of
G. Then p − q + r = 2.
Proof. We use induction on q and Theorem (1.1). For q = 0 the result holds
since we have p = 1 and r = 1, where r is an infinite region. Assume the result holds
for all connected plane graphs with less than q edges and examine G with q edges.
This is where trees come handy - first observe that if G is a tree then p = q + 1
and the formula holds. If G is not a tree then it contains a cycle with some e being
an edge of this cycle. Let us remove e and consider G − e, a connected graph on p
vertices with q − 1 edges and r − 1 regions. The equality p − (q − 1) + (r − 1) = 2
holds by the inductive hypothesis, hence p − q + r = 2.
The two results presented above provide a playground for problems dealing with
restriction on the degree of vertices in a planar graph. In a planar graph G what is
the restriction on δ, the minimal vertex degree?
Corollary 1.1. Every simple planar graph G contains a vertex of degree at most 5.
Proof. Combining equation (1.4) with Euler’s Formula yiels that in a planar
graph G we have q ≤ 3p − 6 which implies that
p
X
deg vi = 2q ≤ 6p − 12
i=1
Chapter 1 Page 8
Recall, that the first equality is given by Lemma (1.1). Suppose that every vertex in
G has degree 6 or more. In this case pi=1 deg(vi ) ≥ 6p, a contradiction. Hence, G
P
Can we come up with a planar graph which contains no vertex of degree less than 5?
An example of such is found in icosahedron, one of five platonic solids. Note that the
exterior of icosahedron should be through of as one of 20 triangular faces and we can
convince ourselves that A and B in Figure 5 are isomorphic 5.
Figure 5. Icosahedron
Bringing our interest in the degree problems one step further, let us prepare grounds
for the final example. In order to solve the problem we need the following side lemma:
Lemma 1.4. Let G be a planar graph in which there is a vertex v0 ∈ V such that
deg(v0 )=2 and every edge e ∈ G is contained in two regions. Then either G = K3 or
there exists a face fi such that fi is not a triangle.
Proof. Let us consider a vertex such that deg(v0 )=2 and its adjecent vertices v1
and v2 . Since every edge is contained in two regions v0 v1 and v0 v2 ⊂ R1 , R2 . If either
R1 or R2 are not triangle we are done. Otherwise both R1 and R2 are bounded by
v0 v1 , v0 v2 and v1 v2 hence G = K3
Example 1.1. (Exercise 7.3 in [2]) Every connected planar graph of order p ≥ 4 has
at least 4 vertices with degree less than or equal to 5.
Proof. The strategy is to obtain a contradiction and show that there does not
exist a graph G of order p ≥ 4 with at most 3 vertices of order less than or equal
to 5. Let p be the number of vertices, p5 number of vertices of degree at most 5, q
number of edges, q1 number of edges which bound only one region, V≥6 vertex subset
5In fact every simple polyhedron can be turned into a planar graph by using the polyhedron’s
vertices as vertices of the graph and the polyhedron’s edges as edges of the graph. The faces of the
resulting planar graph then correspond to the faces of the polyhedron
Chapter 1 Page 9
containing vertices of degree at least 6, V≤5 vertex subset containing vertices of degree
at most 5.
P
Assume that |V≤5 | ≤ 3. Recall Lemma (1.1) to find a lower boundary for v∈V deg(v):
X X X
2q = deg(v) = deg(v) + deg(v) ≥ 6(p − p5 ) + 2(p5 ) (1.5)
v∈V v∈V≥6 v∈V≤5
At this point a careful reader should wonder why vertices of degree 1 were not included
in the above boundary. This is because we are only interested in connected planar
graphs where vertices of degree 1 do not occur. Second, observe that the boundary of
any face f of G is the union of the edges e of G such that e ⊂ f and e ∈ E2 . Let us
define Ω, the set of pairs consisting of a face R and an edge e lying on the boundary
of R as Ω = {(R, e) | e ⊂ R, e ∈ E2 }. Then
X
|Ω| = {e ∈ E2 : e ⊂ R} ≥ 3r
R
X
|Ω| = e = 2(q − q1 )
e∈E2
Therefore
3r ≤ 2(q − q1 ) (1.6)
Where the inequality (1.6) becomes strict if at least one region of G is not a triangle.
Combining Euler’s Formula with inequality (1.6) and equation (1.5) gives:
2 2 1 2
2 = p − q + r ≤ p − q + q − q1 = p − q − q1
3 3 3 3
1 2 2 2
≤ p − ((p − p5 ) + (p5 )) − q1 = p5 − q1 (1.7)
3 3 3 3
Our strategy is to show the impossibility of the above inequality to hold. Since we
assume that p5 ≤ 3 it follows that the inequality can hold only if p5 = 3 and q1 = 0.
Importantly, deg(v) ≥ 2 for all v ∈ V . Now we have two cases to consider
Case 1: There exists v0 ∈ V≤5 such that deg(v0 ) > 2. In this case inequality (1.5)
becomes strict hence (1.7) becomes a strict inequality and we arrive at a contradiction.
Case 2: For every v0 ∈ V≤5 deg(v0 ) = 2. Since q1 = 0 we can apply Lemma
(1.4) and conclude that since p ≥ 4 and so G 6= K3 , there exists a face fi which is
not a triangle. In this case (1.6) becomes a strict inequality and the contradiction
follows.
CHAPTER 2
Graph Colouring
Now that we defined a graph G let us assume that we want to colour its vertices using
colour set C = {c1 , . . . , cn } where each element ci is referred to as a colour. By vertex
colouring we mean a map θ : V → C such that θ(vx ) 6= θ(vy ) whenever vx and vy
are adjecent. A natural question arises of what is the minimal number of colours we
need to use. Mathematics does not like waste - a valid colouring of Cn can surely be
obtained by the means of n colours but in fact the maximum of 3 colours will always
do.
Definition 2.1. The smallest number of colours needed to colour the vertices of G
is called the chromatic number of G and is denoted χ(G). Furthermore, we say that
G is critically n-chromatic if χ(G) = n and χ(G − v) = n − 1 for all v ∈ V (G).
In some cases the chromatic number can be easily established, for instance χ(Kn ) = n,
χ(Cn ) = 2 if n is even and χ(Cn ) = 3 if n is odd and Ku,v = 2. Connected to problems
of colouring is the notion of an independent set, which is a set of vertices of which no
pair is adjacent. The independence number α(G) of a graph G is the size of the largest
independent set of G. Since in a colouring of G all colour classes are independent we
immediately get the following result
χ(G)α(G) ≥ n (2.1)
Theorem 2.1. For every critically n-chromatic graph G the following holds
κ(G) ≥ n − 1 (2.2)
Moreover, Lemma (1.3) and Theorem (2.1) give that if G is critically n-chromatic then
δ(G) ≥ n − 1. We are now ready to present a useful bound for chromatic number
χ(G) due to Szekeres and Wilf as stated in [2].
Theorem 2.2. For any graph G we have χ(G) ≤ 1 + max δ(G0 ) where the maximum
is taken over all induced subgraphs G0 of G.
Thus
χ(G) − 1 = χ(H) − 1 = κ(H) ≤ δ(H) ≤ max δ(G0 )
Which gives that
χ(G) ≤ 1 + max δ(G0 )
Finally, let us offer solutions to two exercises, first found in assigment problems of
Jacob Fox’s notes on combinatorics [5] and second coming from Problem Set 15.1 in
[2].
Chapter 2 Page 12
Example 2.1. Let G1 and G2 be two graphs on the same vertex set V . Show that
the chromatic number of their union G1 ∪ G2 (we take the union of the edge sets of
both graphs) satisfies
χ(G1 ∪ G2 ) ≤ χ(G1 )χ(G2 )
Example 2.2. Show that for any graph G on p vertices χ(G) + χ(Ḡ) ≤ p + 1, where
Ḡ is a graph on the same vertex set as G where two vertices are adjecent in Ḡ if and
only if they are not adjecent in G. Thus, V (G) = V (Ḡ) and G ∪ Ḡ = Kp .
Proof. Choose H, a subgraph of G with max δ(H) and denote maximal δ(H)
as m. It follows from Theorem (2.2)that χ(G) ≤ 1 + m.
Let us find a similar bound for χ(Ḡ). Since δ(H) denotes the smallest vertex degree
within H it follows that |V (H)| ≥ m + 1. Also, for all v ∈ H we have degH (v) ≥ m
where by degH (v) we mean the degree of v in H. Since H is a subgraph of G,
degG (v) ≥ degH (v) ≥ m and from the definition of Ḡ the degree of v as a vertice
of Ḡ is equal to degḠ (v) = p − degG (v) − 1 ≤ p − m − 1. Consider K, any induced
subgraph of Ḡ. We distinguish between the case that K and H are disjoint and the
case that K and H have a vertex in common.
Case 1: K ∩ H = ∅.
Since no vertex of H is shared with K we have |V (K)| ≤ p − |V (H)| ≤ p − m − 1
which implies degK (u) ≤ p − m − 2 for any u ∈ K and hence δ(K) ≤ p − m − 2. So
χ(Ḡ) ≤ p − m − 1 by Theorem (2.2)
Case 2: There exists a vertex v ∈ K ∩ H
Chapter 2 Page 13
1. Map colouring
In this section we will investigate the least number of colours needed for colouring
of any map in a plane. The map colouring problem has notable history on its own
and over the decades it grown to be one of the well known problems in mathematics.
The statement that in order to colour any map one needs at most 4 colours, is
usually attributed to Guthrie, who was to be the first mathematican to conjecture
the so called 4 Colour Theorem in 1852. What initially seemed like an easy problem,
turned out to unsolvale for a long time, despite numerous attempts by distingushed
mathematicians. The proof of the 4-colour conjecture was eventually announced by
Appel and Haken in 1977 and to some disliking of mathematical community it was
the first major proof aided by a computer.
To relate map colouring with graph theory we first need to define a dual G∗ of graph
G and then show that region colouring of G is equavalent to vertex coloruing of the
dual graph of G.
Definition 2.3. Let G be a connected, planar graph. Its dual G∗ is a pseudograph
which has a vertex for each region of G. Two vertices in G∗ are connected by an
edge if and only if the corresponding ragions in G have an edge in common. Dual
Figure 1. For each face of G, the dual graph G∗ (drawn with dashed
lines and empty nodes) has a vertex. Two vertices of G∗ are adjecent
if and only if the corresponsing faces of G are adjecent. Graph H is the
underlying graph of G
From this point it seems like problems of region colouring are equivalent to vertice
colouring. Let us prove it
Proof. Every planer graph has a minimum vertex degree equal to at most 5 as
we shown in Corollary (1.1) Every subgraph of a planar graph is planar and hence
also has a minimum vertex degree of at most 5. Applying Theorem (2.2) we obtain
that χ(G) ≤ 6.
of order p − 1. If deg(v) < 5 then some colour ci has not been used in colouring
of the vertices adjecent to v and by assigning it to v we arrive at 5-colouring of G.
Assume deg(v) = 5 and v1 , . . . , v5 are the vertices adjectent to v assigned distinct
colours c1 , . . . , c5 respectively. Consider H1,3 , a subgraph of G0 induced by all vertices
coloured c1 or c3 . If the vertices v1 and v3 do not belong to the same component of
H1,3 we can interchange 1 the colours 1 and 3 in the component containing v1 and in
such a way obtain a spare colour which we then assign to v. Assume that v1 and v3
belong to the same component H1,3 . Then there exists a path v1 − v3 in G0 . Now
restore the deleted vertex v and note that v − v1 − v3 produces a cycle which closes
v2 . Therefore v2 and v4 belong to different components of a subgraph H2,4 induced by
all the vertices coloured c2 and c4 so by interchanging the colours in the component
containing v2 we gain a spare colour c2 with which we then colour v. Hence we arrived
at valid 5-clouring of G.
1Why can we do it? Since we work on induced subgraph which contains all the vertices coloured
c1 and c3 an interchange does not alter the valid colouring of G0 . All that could go wrong with
changing vertex colouring is possible ending up with at least two adjecent vertices assigned the
same colour. This cannot happen with the above interchange because we do not alter neighbours of
vertices coloured c1 and c3 and by swaping the colour classes we make sure they remain coloured in
a valid way.
CHAPTER 3
(
1 if {vi , vj } ∈ E(G)
aij =
0 otherwise
Theorem 3.1. Let G be a graph with adjecency matrix A = [aij ]. The (i, j) entry of
Ak matrix is equal to the number of different walks of length k from vi to vj .
Proof. For k = 1 the result holds from the definition of adjecency matrix because
a walk of length 1 is just an edge hence if vi and vj are adjecent then aij = 1. Notice
that by the properties of matrix multiplication in the matrix A2 the a2ij entry is given
by X
a2ij = ai1 a1j + ai2 a2j + . . . + aik akj = aik akj
k
Let A n−1
= [an−1
ij ]
and assume an−1
ij is the number of different vi − vj walks of length
n n
n − 1 in G. Let A = [aij ]. By the properties of matrix multiplication we have
X
anij = an−1
ik akj
k
So far we came across some pretty straightforward parameters of graphs such as degree
of vartex, chromatic number, connectivity or girth. As pointed by [5] eigenvalues of
adjecency matrix A(G), although less obvious, are very useful parameter of graphs.
In this chapter we aim to present context in which eigenvalues are used.
The vector x is called eigenvector and the set of eigenvalues is called the spectrum of
A.
Before venturing further into algebraic graph theory let us recall some highlights of
linear algebra which will be used later.
Pn Pn
Theorem 3.2. The sum of eigenvalues is equal to i=1 λi = T r(A) = i=1 aii .
The fact that the adjecency matrix of a graph G has all diagonal entries equal to zero
gives the following
To gain a better understading of algebraic properties of graphs and see the eigenvalues
in action, we set out to investigate the discuss the eigenvalues of a bipartite graph.
Interestingly, the adjecency matrix of a bipartite graph has always the following form:
!
0 BT
AKn,m =
B 0
How does this property translate in terms of eigenvalues? In order to find properties
of bipartite graphs we will proceed with a number of smaller steps. It is convenient
to start by considering a complete bipartite graphs Kn,m .
Lemma 3.2. Let Kn,m be a complete bipartite graph. The only non-zero eigenvalues
√
λ1 , λn are ± mn.
Chapter 3 Page 18
v1 v2
0 0 1 0 0
0 0 1 1 1
1 1 0 0 0
0 1 0 0 0
v3 v4 v5 0 1 0 0 0
Proof. By the remark on the general form of adjecency matrix the rank of Kn,m
is 2. Thus, by Theorem (3.4) we know that Kn,m has 2 non-zero eigenvalues and 0-
eigenvalues of multiplicity n−2. Furthermore, Lemma (3.1) implies that the non-zero
eigenvalues are equal to ±λ, because the sum of eigenvalues must be equal to 0. To
find λ we need to solve Ax = λx:
x1 y1 = xn+1 + . . . + xn+m x1
.. .
..
...
.
x y =x
n+1 + . . . + xn+m
x
n n n
AKn,m = = λ (3.2)
xn+1 yn+1 = x1 + . . . + xn xn+1
.. .. ..
. . .
xn+m yn+m = x1 + . . . + xn xn+m
xT = (α, . . . , α β, . . . , β )
| {z } | {z }
n m
and (Ax)T = (mβ, . . . , mβ, nα, . . . , nα). What is left is solving two equations
mβ = λα
nα = λβ
which leads to
mnβ = λ2 β
√
λ = mn
Chapter 3 Page 19
The drawback of the above result (given by Fox in [5] is in that it holds only for
complete bipartite graphs Kn,m . Ideally, we would like to learn more about a general
bipartitie graph. The next 5 Lemmas and Theorem (3.5) were a part of Jacob Fox’s
Assigment Set [5].
The first equality follows straight from the definition of a closed walk. However, there
is some work to be done to show the second equality. Suppose that λ is an eigenvalue
of A, which means that λ satisfies Ax = λx. Assume λk−1 is an eigenvalue of Ak−1 ,
then Ak−1 x = λk−1 x. Multiplying both sides by A yields
Ak x = A(λk−1 x)
Ak x = λk−1 (Ax)
Ak x = λk−1 (λx)
Ak = λk x
λ1 ≥ −λn (3.3)
Proof. Proof by contradiction. Assume that |λn | > λ1 and choose an odd m.
By the Lemma (3.3)
n n
X X λi
W (m) = T r(Am ) = λm
i = |λn |
m
i=1 i=1
|λn |m
In order to safeguard ourselves from the possible equalities let λn = λn−1 = . . . =
λn−k+1 . Evaluating limit of W (m) as m tends to infinity gives
n−k
X
m λi m
lim W (m) = lim |λn | ( ) + (−1)m k = −∞
m→∞ m→∞
i=1
|λn |
Chapter 3 Page 20
for limm→∞ ( |λλni | )m = 0. Hence, following the |λn | > λ1 assumption we proved that
fom m large enough W (m) < 0. However, the number of closed walks cannot be a
negative integer, hence a contradiction.
The spectrum of adjecency matrix is scattered around zero and there exists a leading
eigenvalue λ1 . Moreover, we know that in terms of absolute values λ1 can be either
leading or equal to the least eigenvalue, λn . Can we say exactly in what cases (3.3)
becomes an equality?
Taking a limit out of (3.4) gives limm→∞ W (m) = ∞, which implies that there exists
some G and some odd m for which W (m) 6= 0, hence it cannot be a bipartite graph.
Conversely, we claim that if G is not bipartite then for all large enough odd numbers m
there is a walk of length m from any vertex vx to any vertex vy . By Perron-Frobenious
Theorem 1 this claim suffices to show that λ1 > λn .
Proposition 3.1. Let G be a connected, non-bipartite graph. Then for all large
enough odd numbers m there is a walk of length m between any pair of vertices.
Proof. Since G is not bipartite by Theorem (1.2) there exists an odd cycle of
length k around some vertex x. For any vertex y ∈ G and y 6= x fix a walk wy0 i from
yi to x and let l1 , . . . , ln−1 be the lengths of wy0 i (see Figure 2). We can choose such
a walk wyi that k divides wyi in the following way:
where w¯y0 i is the inverted yi − x walk and wyi = li k. Take L = max li and set
m = (2L + 1)k. Now what is left is to show that there exists a walk of length m
1Wikipedia [9] gives the following formulation of the Perron Frobenius Theorem: Let A be a real
n × n matrix with strictly positive entries aij > 0, then there is a positive real number r, called the
PerronFrobenius eigenvalue, such that r is an eigenvalue of A and any other eigenvalue λ is strictly
smaller than r in absolute value, that is |λ| < r
Chapter 3 Page 21
w'y
odd cycle
1 y1
of length k
x
w'y
2
y
2
between any two vertices of G. If ya and yb are two distinct vertices different from x,
define a walk yi − yj as follows:
lk
i loop 2L+1−li −lj lj k
yj −→ x −−−−−−−−−−→ x −→ yj
times
By Perron Frobenius applied with Proposition (3.1) and Lemma (3.4) we see that if
G is not bipartite then λ1 > |λn | which completes the proof.
In the next results we will use previous finding to explore the notion of eigenvalue
as a limit of W (m)1/m . Also, we will show that the condition for the graph to be
bipartite is that this limit does not exist.
1
Lemma 3.6. limm→∞ W (2m) 2m = λ1 , where m ranges over positive integers
n
1 1 1
X
W (2m) 2m = T r(2m) 2m = (λ2m
i )
2m
i=1
1
! 2m
n 2m
X λi
= λ2m
1
i=1
λ1
1
n 2m ! 2m
X λi
= λ1 k+
i=k+1
λ1
1
= λ1 k 2m
1
Evaluating the limit of W (2m) 2m as m tends to infinity yields
1
lim W (2m) = lim λ1 k 2m = λ1 (3.5)
m→∞ m→∞
1
Lemma 3.7. If G is bipartite then limm→∞ W (m) m does not exists.
1
Proof. Intuitively, W (m) m will oscillate between positive and negative limit,
depending on the parity of m. Assume that G is bipartite, then as previously shown
n n m ! m1
1 1
X 1
X λi
W (m) m = T r(m) m = (λm
i )
m = λm
1
i=1 i=1
λ1
Now use that G is bipartite if and only if λ1 = −λ1 to show where the problem is
n m ! m1 (
X λ i λn if m is even
lim (−λn )m =
m→∞
i=1
λ1 −λn if m is odd
1
Thus if G is bipartite, limm→∞ W (m) m does not exist.
Next, let us turn to the bounding problem. Although eigenvalues are pretty not-
straight forward features of graphs we can bound them with more feasible parameters
such as minimum and maximum degree.
Theorem 3.5. If G is a connected graph with the minimum degree δ and maximum
degree ∆ then
δ ≤ λ1 ≤ ∆ (3.6)
Chapter 3 Page 23
T
Proof. For the first inequality use the fact that λ1 = max v|v|Av
2 . If we denote the
Finally, we combine the above results to obtain a useful upper bound on chromatic
number χ(G) of G.
where the RHS is equal to the average of the row sums of the matrix A(G’). Since
δ(G0 ) ≥ n − 1, the minimum row of A(G0 ) must be at least n − 1 and hence
λ1 (G0 ) ≥ n − 1 (3.12)
The above upper bound on the chromatic number χ(G) was obtained by Wilf. Impor-
tantly,it is an improvement of Brook’s upper bound χ(G) ≤ ∆ + 1 obtained through
the application of the greedy algorithm. This follows directly from Theorem (3.5) in
which we have shown that λ1 ≤ ∆.
Last but not least let us offer a proof of a well-known Friendship Theorem as im-
mortalized in [1]. In its simplest version the statement goes as follows: In a group
of people where any pair of people have precisely one common friend there is always
a person (also referred to as a politician) who is a friend of everybody. Since the
statement was initially introduced as combinatorial problem, much efforts were put
into conducting a proof relying on combinatorial argument only. Interestingly, the
statement can be shown to be true using an argument based partly on combinatorial
arguement and partly on an application of algebra. The proof presented below is due
to Erdős, Rényi and Sós.
2. Friendship Theorem
If we translate friendship theorem into the language of graph theory we arrive at the
following statement
The proof is by contradiction. Assume that G satisfies the conditions of the theorem
but there does not exist any vertex v which is connected with all other vertices.
First we are going to show that the condition of sharing precisely one neighbour by
each pair of vertices forces G to be a k − regular graph. Note that the condition
of the theorem requires that there are no cycles of length 4 in G. We will call this
requirement a C4 condition.
Chapter 3 Page 25
Lemma 3.8. If G is finite graph in which any two vertices have precisely one common
neighbour then G has to be regular of degree k ≥ 4.
Proof. We choose two arbitrary non-adjecent vertices u and v and let µ(u) and
µ(v) denote the sets of neighbours of u and v. Let us construct a function ϕ : µ(u) →
µ(v) as follows: for any w ∈ µ(u) let ϕ(w) be the unique friend of w and v. Note that
the function is well-defined since v ∈
/ µ(u) so v 6= w. Suppose ϕ is not injective which
means that there exist w1 , w2 ∈ µ(u) such that ϕ(w1 ) = ϕ(w2 ) = z where z ∈ µ(v).
We have z 6= u becasue u ∈ / µ(v) and z ∈ mu(v). Next, z 6= w1 becasue if z = w1 then
w1 = ϕ(w1 ) ∈ µ(w)1 ) which is impossible, similar argument applies to w2 . Hence,
z, w1 , w2 and u are distinct and we have the following situation occuring:
w1
v z u
w2
With this result we progress to algebraic part of the proof which will use Therorem
(3.1) and results on eigenvalues. Fix v and look at walks of length 2 from v. We can
distinguish two kinds of such walks:
Case 1: The walk is closed. In such a walk we basically go forth to one of the
neighbours of v and then back to v. There are deg(v) = k such walks
Case 2: The walk is not closed. Note that the walk is completely determined
by its endpoint u (because v and u have a unique common neighbour) which implies
that for all u 6= v there exists unique walk of length 2 which terminates at u hence
there exists n − 1 such walks.
Summing over the degrees of the k neighbours of v we get
W (2) = k 2 = k + n − 1
where I is the identity matrix and J is the matrix with all entries equal to one. Next,
we would like to find eigenvalues of matrix A2 and A. The rank of J is 1, so there
is one nonzero eigenvalue equal to n, similarily the rank of I is n and 1 is the only
eigenvalue with multiplicity n. It follows that A2 has eigenvalues k − 1 + n = k 2 of
multiplicity 1 and k − 1 of multiplicity n − 1. This implies that A has eigenvalues k or
√ √
multiplicity 1 and ± k − 1. Suppose that r of the eigenvalues are equal to k − 1
√
and s of eigenvalues are equal to − k − 1 where r + s = n − 1. Since the sum of
eigenvalues is equal to zero by Theorem (3.1) we have
√ √
k+r k−1−s k−1=0
since r 6= s
√ k
k−1=
s−r
√
At this point we need to use Dedakin’s result which says that if a square root m of
√
natural number m is rational, then it is an integer. Let h = k − 1 ∈ N then
h(s − r) = k = h2 + 1 (3.14)
Now since h divides h2 + 1 and obviously h divides h2 , we conclude that h can only be
equal to 1 and therefore k = 2 which we ruled out, hence a contradiction is obtained.
CHAPTER 4
In the following two chapters we will explore the areas where graph theory intersects
probability theory. Our aim is to introduce the notion of a random graph. Next,
in the context of probabilistic method we will to show that there are graphs with
arbitrarily large girth and chromatic number.
1. Random graph
A random graph is a graph generated by some random process. To be more precise let
G be a random graph on n vertices and let us flip a coin which comes up heads with
probability p independently for each of the n2 potential edges. If the coin comes
up head then the edge e ∈ E(G), otherwise the edge e ∈ / E(G). The probability
space G(n, p) consists of all graphs on n vertices where the individual edge e appears
with probability p, independently from each other. For instance, the probability of a
n
complete graph Kn would be P (Kn ) = p(k ) , since all the possible edges would need
n
to belong to G. Generally, we have P (G) = pn (1 − p)( 2 )−m if there are exactly m
edges in E(G).
2. Erdős’ Proof
Recall that girth γ(G) is the length of the shortest cycle contained in the graph G and
chromatic number χ(G) is the smallest number of colours needed to colour a graph G
so that no two adjecent vertices are assigned the same colour. Observe that choosing
a graph with high chormatic number or high girth alone does not pose a probelm
and is the question of existance of arbitrarily high chromatic number and high girth
that is interesting. For example in Kn , a complete graph on n vertices, we have
χ(Kn ) = n, but γ(Gn ) = 3. On the other hand in Cn , a cycle graph on n vertices, we
have γ(Cn ) = n, but χ(Cn ) ≤ 3. Intuitively, it may seem that high chromatic number
goes at the price of low girth and vice versa. The assumption that we can either
achieve high chromatic number or high girth seems plausible. Surprisingly, such an
intuition in this case is wrong and a proof due to Erdős presented below shows that
there exists a graph G with arbitrarily high chromatic number and girth.
27
Chapter 4 Page 28
The elegance of this method lies in the fact that it allows to prove existence of difficult
objects (such as a graph in question), without creating or providing a method of
constructing the object directly. Indeed, anyone who tried to give an example of
graph with high girth and high chromatic number will find that this is not an easy
task. In the light of this discussion the following theorem comes as a surprise.
Theorem 4.1. For every k ≥ 2 there exists a graph G with χ(G) > k and γ(G) > k.
By the probabilistic method, in a certain probability space G(n, p) on all graphs with
n vertices the probability that a graph has either χ(G) ≤ k or γ(G) ≤ k is strictly less
than 1, therefore there exists a graph G ∈ G(n, p) with χ(G) > k and γ(G) > k. The
strategy is to show that on a certain probability space, which we will choose wisely,
the probability for χ(G) ≤ k is smaller than 21 and the probability for γ(G) ≤ k is
smaller than 12 . Hence there exist a graph G with desired properties.
2.1. Chromatic number. We start by noting that vertices which belong to the
same colour class must by definition be pairwise non-adjecent and hence they form
an indepenent set.
Chapter 4 Page 29
[ X
P (G : α(G) ≥ r) = P ( CR ) ≤ P (CR )
R⊆V R⊆V
n r r
= (1 − p)(2) ≤ nr (1 − p)(2)
r
r−1 −p(r−1)
= (n(1 − p) 2 )r ≤ (ne 2 )r (4.1)
n 1
P ({G : α(G) ≥ }) < (4.2)
2k 2
What comes next may seem surprising but in fact it is nothing more than using our
special set-up to validate (4.2). Firstly, we have
1
n k+1 ≥ 6k ln n (4.3)
1
for all n large enough. We can divide both sides by n > 0 to obtain
1
n k+1 1 k+1 −k 6k ln n
= n k+1 − k+1 = n k+1 = p ≥
n n
n
For r := 2k
this gives
pr ≥ 3 ln n (4.4)
Having (4.3) in mind we continue evaluating (4.1) as follows
√
−
p(r−1)
− pr + p2 − 32 ln n 1
− 12 1 e
ne 2 = ne 2 ≤ ne e =n
2 e = √ → 0 as n → ∞
2 (4.5)
n
1To see this use l’Hôpital’s rule, which gives
6k ln n (6k ln n)0 6kn−1
lim 1 = lim 1 = lim k =0
1 − k+1
n→∞ n k+1 n→∞ (n k+1 )0 n→∞
k+1 n
Chapter 4 Page 30
p 1
Note that e 2 ≤ e 2 since p ≤ 1. Hence, (4.2) holds for all n ≥ n1 for some large n1 .
2.2. Girth. Let us turn to the second parameter, girth γ(G). For a given k our
aim is not to have too many short cycles of length less or equal to k. Let 3 ≤ i ≤ k
and consider a fixed i-set I ⊆ V . The number of possible i-cycles on I is equal to
(i−1)!
2
, where the division by 2 accounts for the fact that each cycle can be travelled
in two directions. The total number of possible i-cycles in V is therefore ni (i−1)!
2
and
every such cycle on I appears with probability pi .
k
Let X be the random variable on G(n, p) where p = n− k+1 , counting the number of
cycles of length less or equal k in a graph G ∈ G(n, p). For every i-cycle I where i ≤ k
with vertices in V , the vertex set of the graphs in G(n, p), let XI : G(n, p) → {0, 1}
be the indicator random variable of I:
(
1 if I ⊆ G
XI : G 7→
0 if I * G
n
2.3. Putting it together. Let an event A = {G : α(G) ≥ 2k } and event
B = {G : γ(G) ≤ k}. Up to this point it has been shown that for all n >> 0
1 1
P (A ∪ B) ≤ P (A) + P (B) < + =1 (4.6)
2 2
Therefore the probability of complimentary event P ((A ∪ B)C ) > 0. Hence there
n
exists a graph H ∈ AC ∩ B C such that α(H) < 2k and H contains less than n2 short
cycles of length less or equal to k. If we delete one vertex from each of the short cycles
then for the resulting graph G we would have γ(G) > k. The question is whether
we can delete vertices without ruining the high chromatic number of H in the same
time. Suprisingly, in this particular set-up the answer is yes. Deleting vertices does
affect the chromatic number but not at the pace that would spoil the final result.
Chapter 4 Page 31
We certainly deleted less than n2 vertices from the graph H so the resulting graph G
contains more than n2 vertices and satisfies α(G) ≤ α(H) < 2kn
. Finally, we arrive at
n/2 n n
χ(G) ≥ ≥ > =k
α(G) 2α(H) n/k
which finishes the proof.
CHAPTER 5
What is percolation?
1. Set-up
Figure 1. Graph of θ.
By application of Kolmogorov’s 1-0 law2, which we will not present here, we have
an interesting situation: if p < pH then the probability of the even E that there is
an infinite open cluster is 0, while if p > pH this probability is 1. What this means
is that the component structure of Λ is quite different for p < pH and p > pH .
Terminology-wise, in a given model we say that percolation occurs if θx > 0 and
P (E) = 1.
Although for p < pH the open cluster Cx is finite, the expected size of Cx does not
to be finite. The second critical probability, pT provides information about when the
expected size of Cx is no longer finite. Let
χx (p) = Ep (|Cx |)
χ(p) is increasing with p and is finite for some site x if and only if it is finite for all
sites. We define pT as follows
In general the critical probabilities introduced in the previous section are not easy
to find. However, in the example to follow we can obtain both pT and pH at an
affordable price.
Theorem 5.1. (k-branching tree) Let Tk be an infinite rooted tree in which each
vertex has k children.
1
pT (Tk ) = pH (Tk ) = (5.1)
k
v0
πn = 1 − (1 − pπn−1 )k (5.2)
Let us define a function fk,p = 1−(1−px)k , then (πn ) = fk,p (πn−1 ). We are interested
in what values fk,p takes on the interval [0, 1]. First observe that we can differentiate
3For example it is believed for the expected number of vertices in Cx χ(p), the following holds:
− ln χ(p)
ln (p−pc ) → γ for a suitable constant γ.
Chapter 5 Page 35
hence fk,p is increasing and concave. Furthermore, fk,p (0) = 0 and fk,p (1) < 1, hence
0
there exists some x0 ∈ (0, 1) such that fk,p (x0 ) = x0 if and only if fk,p = kp > 1.
f'k,p(0) = kp > 1
)
(x
f k,p
y=
x0
1
Bearing in mind that (πn ) = fk,p (πn−1 ) we see from the graph that if p > k
and
x > x0 then x0 < fk,p (x) < x. Since π0 = 1 we have that
implying that
1
pH (Tk ) ≤
k
1
On the other hand if p < k
then πn converges to 0 and θv0 (Tk ; p) = 0. The critical
probability pT must be k1 .
For the pT , observe that the probability that a site y which is at distance l from vo
belongs to Cv0 is equal pl and therefore
Chapter 5 Page 36
∞
X
χv0 (Tk ; p) = E(|Cv0 |) = k l pl (5.3)
l=0
1
where (5.3) if finite for p < k
and infinite for p ≥ k1 . Therfore, the critical probability
pT (Tk ) = k1
Next, we are going to discuss the bounds on the critical probabilities pH , pT on square
lattice Z2 . Recall that the dual Λ∗ of a graph Λ drawn in the plane has a vertex for
each face of Λ and an edge e∗ for each e. An edge e∗ ∈ E(Λ∗ ) joins two vertices of Λ∗
which correspond to two faces of Λ in whose boundary edge e ∈ E(Λ) lies. In studying
Z2 the convention is to take v∗ = (x + 12 , y + 12 ) as the dual vertex which correspons to
face bounded by following vertices:(x, y), (x + 1, y), (x + 1, y + 1), (x, y + 1). The sites
and bonds of Λ∗ are called dual sites and dual bonds, with the dual bond e∗ open if
e is closed and vice versa.
Before turning out attention to the critical probabilities on square lattice, one more
result which will be used in the boundary proof is needed.
Definition 5.1. A self-avoiding walk µn (Λ; x) on a square lattice is a path which
never intersects itself.
The above inequality comes from the fact that the number of self-avoiding walks
must be smaller than a number of the so called nonreversing walks, which are ran-
dom walks whose alternating sequence does not have any subsequence of the form
vn , en , vn+1 , en , vn . Hence, for the nonrevesing walk there are 4 choices for the firts
step and three choices for each step therafter.
If H is a finite, connected subgraph of Λ with vertex set C, it follows that there
exists an infinite component C∞ of Λ − C, which is the subgraph of Λ induced by the
vertices outside H. Let σ ∞ C be the set of bonds of Λ∗ dual to bonds of Λ joining C
and C∞ .
Proposition 5.1. If C is the vertex set of a finite connected subgraph H of Λ = Z 2 ,
then σ ∞ C is a cycle with C in its interior, hence it is a boundary.
Chapter 5 Page 37
We will not give a detailed proof but outline the technique of the prove given in [3].
→
− →
− →
− → −
Denote F to be the set of the C−C∞ bonds, oriented from C to C∞ . For f = ab ∈ F
→
− → so that a is on its left. Let −−−→ → − → − → −
orient to dual bond f ∗ = −uw σ ∞ C = f ∗ : f ∈ F . Then
→
− →∈− −−→
through a straight-forward consideration it can be shown that if f ∗ = − uv σ ∞ C then
−−−→
there is unique bond of σ ∞ C leaving v, which proves that the underlying unoriented
graph σ ∞ C consist of a single cycle S, with C∞ being outside of S and C inside S.
Since any dual cycle S must cross the x-axis at some coordinate x∗0 such that k + 21 ≤
x∗0 ≤ 2l−3
2
, there is fewer then l choices for e∗ . If we split S into e∗ and the remaining
path of length 2l − 1 it becomes clear that there are at most lµ2l−1 choices for S.
Denote the number of open dual cycles surrounding Lk by Yk . For any closed cycles
in Λ∗ surrounding Lk let zS be the random variable such that
(
1 if Λ∗ contains S
zs (G) =
0 otherwise
And Yk (G) = S zS (G) = S⊆Λ∗ 1. Keeping in mind that a dual edge e∗ ∈ E(Λ∗ ) is
P P
[1] Aigner, M., Ziegler, G. 2004.Proofs from The Book. Berlin: Springer
[2] Behzad, M., Chartrand, G. 1972.Introduction to the theory of graphs. Boston: Allyn and Bacon
[3] Bollobas, B. Riordan, O. 2006. Percolation. Cambridge : Cambridge University Press
[4] Cameron, P. 1994. Combinatorics : topics, techniques, algorithms. Cambridge : Cambridge
University Press
[5] Fox, J., MAT 307 Combinatorics (Spring 2009) [Course notes, available from
http://www.princeton.edu/ jacobfox/MAT307.html Accessed on February 02, 2010]
[6] Grunbaum, B., Soifer, A., Rousseau, C. 2008. The Mathematical Coloring Book: Mathematics
of Coloring and the Colorful Life of Its Creators. New York: Springer
[7] Kesten, H.What is...Percolation?. Notices of the American Mathematical Society. Vol. 53 No
5. August 2006
[8] Pegg, E. Snark. MathWorld. Available at http://mathworld.wolfram.com/Snark.html Ac-
cessed on February 10, 2010
[9] Wikipedia: The free encyclopedia. Wikimedia Foundation, Inc. Available at
http://www.wikipedia.org Accessed on February 02, 2010
39